More companies are looking into model post-training, but very few people have the skills. Demand is growing.
post-training is seen as a rare skill partly because many ML engineers who entered the field in the last 3–4 years have never actually trained a model, let alone understood its architecture.
And that's fine. Most never needed to.
The real scarcity comes from how few labs had the resources to train large models.
At its core, post-training is training at scale on specialised data.
If you're hiring for it, don't just look for people with "LLM post-training" on their CV.
Look for people who've done deep ML research, trained large models, and genuinely understand the architecture and the data.
The skills are more transferable than the job titles suggest.
41
What happens when a country teaches 200 million kids to learn with AI?
China is finding out, and the early numbers aren't what it hoped.
A new Reuters investigation, based on rare school visits plus 100+ studies and 100+ government documents, shows how far it already goes:
AI agents writing personalised lessons for 4,000 pupils in one school.
Kids building apps and robots
One region using AI to record and analyse classrooms, including student behaviour.
The early numbers: in a study of about 27,000 teenagers, AI use went with 18% higher homework scores, but 20% lower exam scores.
And a survey of 320,000 pupils found frequent users scored lower on memory, comprehension, analysis and creativity.
China is betting the whole system.
If it works, it gets a generation that thinks with AI.
And If it doesn't, it gets 200 million kids who learned to hand in homework, not to think..
Going to link to an article in the comments
Why would investors pay 100x revenue for an AI startup?
Harvey raised at about 50x its revenue. Sierra at 100x. These are real rounds, and if you're an AI founder raising now, this is what investors are comparing you to.
The reason is speed. The fastest AI startups reach $100M in annual revenue in about 18 months. The best software companies before AI took around 7 years. And the top AI companies are still growing 115% a year after they pass $100M.
It shows up early too.
The median Series A for an AI model startup in Q1 2026 was about $300M.
For a non AI startup it was about $55M. Investors aren't paying for what you've built.
They're paying for how fast they think you'll grow..
2
43
United just banned personal AI agents from booking flights for you.
Their reasons: shared passwords can leak, an agent might spend your miles, and an agent checking you in could even break safety rules.
I don’t read it as anti-AI. It’s an airline being an airline. Safety first, and no one wants the call that starts with “it wasn’t me, it was my agent.” Who’s responsible for a wrong booking? The customer, the agent, or the airline?
Until that’s answered, the safe answer is no. Then they’ll work out the model: their own agent, approved partners, or something nobody’s built yet.
Would you trust an AI agent to book your flights before the airline does?
55
Jensen Huang, CEO of NVIDIA, has announced his immediate resignation.
NVIDIA CFO Colette Kress will take over as interim CEO.
56
They promised AI would save us time (and it does).
But somehow, for now, I mostly spend that time convincing it that it's wrong. At least with humans you can give up and move on..
17
OpenAI just bought a 9-person startup that ran almost entirely on agents. The founder’s last company, Sonder, needed over 1,000 people to reach $600M in revenue.
Odessia is an AI travel agent, less than a year old, backed by Sequoia five months ago. The interesting part is how it was run. Davidson says almost every job started with an agent: strategy, marketing, customer support, legal, finance.
Tasks that took 40 hours were done in about two.Repeated corrections were turned into reusable skills. Some agents worked with no one starting them: a daily analytics report, an investigation into every production error, smoke tests overnight.
The team’s day began by reviewing what the agents had done while they slept.The bottleneck is no longer doing the work.
It’s how fast a person can review it and decide.
If your day started with reviewing what agents did overnight, what would you hand over first?
31
Almost science fiction: Norway is protecting its salmon with AI lasers.
A company called Stingray has built over 3,000 underwater units for fish farms.
Cameras watch every fish that swims past, the AI confirms it has spotted a sea louse, and a green laser kills it.
The fish keeps swimming.
The old fix was pulling the fish out and blasting them with hot water or high pressure.
Norway's Veterinary Institute says injuries from those treatments are now the leading cause of death in farmed salmon. 54.9 million died in Norwegian sea cages last year.
The laser isn't the impressive part. The AI is: it makes a one-at-a-time job possible across tens of millions of fish.
Work that was too fiddly to do individually suddenly becomes cheap.
What other "one at a time" job is AI about to make normal?
44
What a crash. ZoomInfo was worth $26bn in 2021. Today it's worth about $1bn.
The CEO explained why on last week's earnings call, quoting a customer: "I need your data, and I need way more of it than I have ever needed before, but I do not need the seats."
AI tools pulling ZoomInfo's data through its API doubled in a quarter. Existing customers pay 12% less a year.
ZoomInfo is now moving from per-seat pricing to credits.
One agent on one API key can do the work of 50 seats. Which per-seat software is next?
I don't want to use your AI assistant on your website. I want to use *my* AI assistant on your website.
2
34
Apple just started charging for AI. $49.99 for an AI top-up in Logic Pro.
It's an assistant in Logic Pro that does the work for you.
Type "set up a vocal track" and it adds the track, opens the plugins and sets the level.
But the catch:
How much do you get? Not stated. When does it run out? 12 months after you buy, used or not.
AI subscriptions are already vague, but this is a one-off purchase with no quantity on it.
Would you pay for a top-up with no size on the label?
1
2
56
Aina’s new AI ring guesses your next three moves and lets you accept with one tap.
Super cool and Impressive tech.
But what happens It’s that tapping yes gets so easy you stop noticing when the reply should have been yours.
Nobody decides to outsource their thinking. It happens one convenient tap at a time.
What do you think?
What if your devices, apps and agents knew what you wanted?
You finish breakfast, and your Uber is booked.
Running late, your agent sends a “Running late” text.
You reach home, and your lights turn on.
AI is proactive now. But it still needs your yes. It waits, for you to unlock your phone, open an app and tap approve. Every time, hundreds of times a day.
AI needs its own interface.
Aina is a ring with a screen that predicts and shows you the three things you’ll need next. One flick of your thumb, and it’s done. For everything.
✦ Cues: proactive suggestions, ready before you need them
✦ Voice: talk to your agents, take notes, dictate
✦ Remote: control your music, Mac and TV
✦ Calls: right from your hand
Pre-orders open now
aina.com
58








